An example where the business case is at odds with the best situation for the user.

The ideal would be a local hub that can easily provide local area "app" connectivity, with ability to work with a DNS provider, Let's encrypt, and the relevant firewalls to let it have it's own port. Regrettably every member of the industry has steered explicitly away from making this interoperable. Upnp was designed to faciltate asking for real ports, but generally considered insecure and disabled with no standardized authenticated way to do the same. We basically didn't create an ecosystem capable, because it being anti-revenue.

The money is on the side of client being a dumb slave to vendor curated internet services that happen to also let them hold your device hostage for a subscription fee.

Home Assistant Green is the closest to enable it, but you get to own having to see if the devices are willing to work with it but without cloud and you get to own the task of getting it a port going to it from the internet, it's a piece in an ecosystem that's barely willing to accommodate the use case.

Just dealt with some nest thermostats. Hate their design.

Much prefer the "Sensi" thermostats, but the ones that still have HomeKit since those can be onboarded onto a LAN without internet (and even without an Apple Device). However they seem to have decided to discontinue HomeKit in their newer models, so I won't touch them either.

Part of it though is those sorts of classes were stupid not because those skills were useless, but because you didn't really *need* the classes as the software was easy to use, unless you got into some of the more tricky spreadsheet stuff that looks more like programming than just office stuff.

Similar here, to the extent LLM *can* augment/implement coding, they really don't need a university major to tell them how to use AI... If it can work, it can work easily. You don't need to be *taught* how to use the LLM, it should come relatively naturally.

Problem being that LLM in university is going to be totally different than LLM in real life. LLM in university/school is going to feel a lot more capable than at least some of the professional world. A university curriculum won't ask you to do something that hasn't already been done, because they need precedent to know if they are making an unreasonable assignment or not. In the real world you are potentially going out on a limb. There's of course a lot of work that is well trodden territory that LLMs might just chug away at, but at least some of the market is beyond that.

I've never had particularly compelling results from prompt style interaction, though as a code completion it has been... ocassionally useful, still usually wrong, but it can generate almost-correct code worthy of fixing faster than I can type it 10-20% of the time.

It could just be the area I develop in, which is a bit more niche than maybe what other people are doing, but it seems to struggle a lot with having no clue about the ecosystem I work with day to day.

I hear you, as a member of the first generation to be cluttered with software developers who jumped in for a gold rush, it may be nice, eventually, for things to settle in.

But we still have a great deal of grift from those gold rush people who can convince managers that they are best because they can use Claude to write up whatever the manager wants.

The problem is that it is incapable of "realizing", so if it's spiraling out, it's really not worth trying to make it correct itself, it's not "learning".

To the extent the chat guides things to a correct path it is by influencing the statistics of the content away from a failing outcome.

It will happily be correct, but admit it is wrong if the human contribution to the text content and then say something wrong, then pivot back if that's what the human says. None of this influences the 'next' person to come along, since that interaction didn't change the model any, unless fed back into the training set later, but that's not necessarily a good way to improve the model.

Maybe a job of nothing but coding would be dead but I don't think such a job should have ever been alive.

I can't imagine anyone who has actually used LLM coding assistance think the skill of being able to read, modify, and write code would be dead anytime soon. It went from complete absurdity to surprisingly capable, but still mostly wrong real quick and has kind of sat there for all but the most brain dead simple projects.

I have one of those knob controls for the phone interface and it's good for most things, but if you want to generally peruse the broader map around your route touching is nicer since you can zoo easily rotate, scroll, and zoom. I frequently do that when I'm setting out, request a route, then fiddle with it a bit to understand the route before actually starting the drive, having it zoom out to show me the big route and then I pinch to zoom in on any interesting looking turns so I know what's coming/decide if I want to do it a slightly different way (google loves to just take an exit and get back on to shave maybe one minute off a traffic jam, and I generally ignore those).

Not just tactile feedback, but ability to feel. If they just vibrated the screen a bit, it's an improvement, but doesn't help with navigation.

There was this one car I was in some time ago that did have hard controls, but it was mostly a fairly large sea of nondescript square buttons. Almost as bad as a touch screen.

Having a control surface you can feel with obvious knobs and such with distinct feels for at least the most prominent functions..

I'm not so sure about this, this has been one area where the 'natural language' interfaces sometimes lose a lot. For example, if you are trying to review a map on your display, just nothing beats being able to fling it around, twist it, and zoom it with your fingers. For a lot of adjustments, it's nice having the full range so you can zip it straight to where you want without having to audibly describe where you want. And of course hard controls for adjustments and instant reactions, like 'shut the music up, answer/hang up a call, adjust the temperature real quick'. Voice for things like 'navigate to work', maybe, though on the other hand Google already presents a selection of two or three likely destinations, and work or home will be one of those so tapping is quicker, and if it got smarter then sometimes it might catch other destinations too.

In other areas I've seen people champion ditching UIs because natural language is here, but there's a lot of stuff that's harder to do with language than to hit buttons or knobs or such.

My car has a 15 inch touchscreen that can just be a projection of my phone's interface fixed in a supremely convenient location and nice and huge.

I'm with you on leaning too much on the built-in hardware which will age like milk compared to the expected lifetime of the car, which is why things like Android Auto and Apple CarPlay are so useful to get the interface benefits of the built in, but with the core software and hardware capable of being modular.

At least as of the 2022 model, that was still alive, just concurrent with the touch interface.

They want a touch screen because you need at least some touch screen to do Apple CarPlay or Android Auto or other similar ecosystem applications.

For example, a touchscreen is nearly a requirement for decent GPS UI. There are applications where the multitouch gestures are hard to beat for the UI.

Problem is the industry struggles with the concept of "touch screen for *some* things, but not everything". So you get things like "here's the touchscreen you wanted, *including* replacing those HVAC hard controls with some touch screen stuff".

Of the modern cars I've been in, a 2022 BMW seemed to get it pretty much just right, a healthy amount of physical buttons and a respectable touch screen along with a wheel with haptic feedback for eyes-free cursor style interaction even with Apps. The had this goofy "air gesture" thing that was pretty useless, but it seems like they've given up on that.

For the missing child scenario, they cast a wide net because they want to try to cover about a 2-3 hour radius of likely travel so that someone might see a license plate. But as was pointed out, this can be useful for people actively driving or in a parking lot, not so useful when it's on a nightstand indoors at home where there's zero chance the person is going to see any license plate.

While his comment was a bit off putting, there's a point about overuse of the system.

A child goes missing. That's unfortunate. Perhaps not as unfortunate as you might guess, it's often a custody dispute with lower risks than the alert would make you think, but still worth getting the word out.

But 150 miles away, on a nightstand where it's certain that I won't be recognizing the license plate in the alert, there's not much to expect that I'll be able to do anything helpful. Sometimes the alerts are pointless. One was something like "we are searching for a six year old girl", no vehicle description, not even a description of any person apart from the age and a gender.

On the weather side, if it even thinks of raining, I'm going to get likely a number of emergency alert sounds with flash flood watch. A very normal rain pattern that could flood certain places to be sure, but those places are used to assuming any rain means a flood, so it's not news to them. Whether it's a "gentle rain will cause normal flooding" or "a catastrophic hurricane is going to level the whole area", it's the same alert with same tone of urgency and a single off/on switch to say whether you get them or not.

It's hard to take them seriously when 90% of them are nothing and/or I'm useless to them, but they present themselves with the same level of critical urgency demanded no matter what.